SoulX-Podcast is an inference codebase for generating high-fidelity, long-form podcasts from text. It targets users needing realistic multi-turn, multi-speaker dialogic speech synthesis, offering advanced features like cross-dialectal zero-shot voice cloning and paralinguistic controls for enhanced naturalness and personalization.

How It Works

The project focuses on realistic long-form podcast generation, excelling in multi-turn, multi-speaker dialogic speech synthesis. It integrates a range of paralinguistic controls (e.g., laughter, sighs) to enhance realism. A key novelty is its support for cross-dialectal, zero-shot voice cloning, enabling personalized speech generation across various Chinese dialects (Sichuanese, Henanese, Cantonese) and Mandarin/English, using prompt audio samples.

Quick Start & Requirements

• Installation: Clone the repo (git clone git@github.com:Soul-AILab/SoulX-Podcast.git), create a Conda environment with Python 3.11 (conda create -n soulxpodcast -y python=3.11), activate it (conda activate soulxpodcast), and install requirements (pip install -r requirements.txt).
• Model Download: Download base and dialectal models (1.7B parameters) via huggingface-cli or Python snapshot_download. Git LFS is required for git clone download.
• Prerequisites: Conda, Python 3.11, Git LFS.
• Usage: Basic inference can be run via bash example/infer_dialogue.sh.
• Links: Demo page: https://soul-ailab.github.io/soulx-podcast/. Paper: https://arxiv.org/pdf/2510.23541. Hugging Face models: https://huggingface.co/collections/Soul-AILab/soulx-podcast.

Highlighted Details

• Generates long-form, multi-turn, multi-speaker dialogic speech.
• Supports cross-dialectal, zero-shot voice cloning for personalized speech.
• Integrates paralinguistic controls like laughter and sighs for enhanced realism.

Maintenance & Community

• Paper published: https://arxiv.org/pdf/2510.23541.
• Models available on Hugging Face: https://huggingface.co/collections/Soul-AILab/soulx-podcast.
• Demo page: https://soul-ailab.github.io/soulx-podcast/.
• Contact emails provided for inquiries.
• A WeChat group is available for technical discussions.
• Future development includes a WebUI, online demo, and Docker support.

Licensing & Compatibility

• License: Apache 2.0.
• Compatibility: Researchers and developers are free to use codes and model weights. The Apache 2.0 license generally permits commercial use and linking with closed-source projects.

Limitations & Caveats

• The project is primarily an inference codebase; example scripts for monologue TTS are pending.
• A WebUI, online demo, and Docker support are planned but not yet implemented.
• Streaming inference is also a future development goal.
• A usage disclaimer strongly advises against misuse for unauthorized voice cloning, impersonation, fraud, or illegal activities, emphasizing ethical standards and responsible AI use.